Accurate runoff simulation is of great importance to understand watershed hydrologic cycle process, effective utilize water resources and respond flood disaster. Hydrologic model is one of the main tools for runoff simulation research and the continuous improvement in Machine Learning offers powerful tools for modeling of hydrologic process. This research took the runoff process of the Atsuma River basin in Hokkaido from 2015 to 2019 as object, proposed a special machine learning framework: Long-and Short-term Time-series Network (LSTNet) for runoff simulation, discussed the accuracy for runoff simulation of LSTNet model with (multivariate LSTNet Model) or without (univariate LSTNet Model) meteorological factors and Soil and Water Assessment Tool (SWAT) model respectively, analyzed the model selection for runoff simulation under different data conditions in the basin. The Nash-Sutcliffe efficiency coefficients (NSE) of the runoff simulation results in the validation (test) period were 0.633 (SWAT model), 0.643 (multivariate LSTNet model), and 0.716 (univariate LSTNet model) respectively. The results show that the accuracies of the two models for runoff simulation in the Atsuma River basin are all very high. SWAT model has prominent advantages in runoff simulation and shortcomings. LSTNet model shows great advantages and potential in runoff simulation. In summary, when target basin’ s data is accurate and complete, the accuracy of SWAT model in runoff simulation is high and stable. When the target basin lacks data or the quality of data is poor, LSTNet model can realize high-precision runoff simulation only based on the measured runoff data, which has a strong application.